nishant-pall / asr

Automatic Speech Recognition

Objective

To use the Deepspeech model from Mozilla, and use it extensively for identifying the software engineering jargon to assist developers in sorting issues and debug their code.

Contents

Resources

• Deepspeech Documentation
• Deepspeech Paper
• KenLM Library
• Scripts obtained from Object-Oriented-Design from the University of Alberta on Coursera, and
• Server-Side Development from The Hong Kong University of Science and Technology.

Formatting

• Back-End Web development text scripts.

• Object Oriented Design text scripts.

• Notebooks for appropriate formatting and work up to make scripts suitable for training a scorer have been named according to their function.

Building Scorer

• Scorers were built here.

• Follow the External Scorer Scripts documentation for building and training your own scorers according to the dataset.

• STEPS TO BUILD A SCORER

  • Gzip the OO.txt to oo.txt.gz using

    gzip OO.txt
  

  • Build the KenLM Library.

  • Use the following command to build the lm.binary and vocab.txt, as done in createVocab.sh.

   python3 generate_lm.py 
   --input_txt oo.txt.gz 
   --output_dir . \
   --top_k 500000 
   --kenlm_bins ../../kenlm/build/bin/ \
   --arpa_order 5 
   --max_arpa_memory "85%" 
   --arpa_prune "0|0|1" \
   --binary_a_bits 255 
   --binary_q_bits 8 
   --binary_type trie
  

  • To build the scorer, run the following command, as done in build.sh
  • Fine tune values of parameters default_alpha and default_beta for obtaining signals appropriately.

   ./generate_scorer_package
   --alphabet ../alphabet.txt
   --lm lm.binary 
   --vocab vocab-500000.txt \ 
   --package kenlm.scorer
   --default_alpha 0.93
   --default_beta 1.18 
   --force_bytes_output_mode True
  

Testing

• Audio file obtained from Object-oriented Programming in 7 minutes | Mosh.

• Used convert.ipynb for extracting the audio from the video and converting it into .wav format.

• To obtain the result script and store it into result.txt run the following command in terminal.

  deepspeech --model deepspeech-0.9.3-models.pbmm --scorer kenlm.scorer --audio OO.wav > result.txt